Weakly supervised rotation target detection method in remote sensing images based on view consistency network

Abstract

The present invention discloses a method for detecting rotation targets in remote sensing images based on a weakly supervised view consistency network, comprising: preprocessing the original satellite image I to obtain a set of horizontal candidate regions {H1, ..., H N}; Use the backbone network to extract the feature map M∈R of the original satellite image I h×w×l , where h, w, l are the length, width and channel dimensions of the feature map respectively; the feature map M∈R h×w×l and the horizontal candidate region set {H1,…,H N A pre-established and trained rotation target detection model is input to obtain a first type of rotation detection frame, and the first type of rotation detection frame is used as the target recognition result. The present invention's view consistency network-based weakly supervised remote sensing image rotation target detection method reduces the workload of manual annotation and improves the detection efficiency of weakly supervised targets.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, in particular to a method based on image processing technology. Figure 1 Weakly supervised remote sensing image rotation target detection method based on consistent network. Background Art

[0002] In recent years, object detection in remote sensing imagery has accelerated the digitalization of dynamic land monitoring and law enforcement using satellite imagery. Due to the bird's-eye view of remote sensing imagery, detected objects are not limited to vertical or horizontal positions but can be arranged at various angles. Consequently, the horizontal detection box used to locate objects in traditional object detection is no longer able to accurately describe the target's position. Therefore, rotating object detection in remote sensing imagery has become a critical research task in the field of vision.

[0003] With the introduction of convolutional neural networks, a lot of deep learning-based rotation target detection work has emerged. Without exception, the training of these models is based on a large amount of labeled data. Taking the two mainstream benchmark datasets in this field, DOTA-v1.0 (a large-scale dataset for target detection in aerial images) and DIOR (a large-scale benchmark dataset for target detection in optical remote sensing images) as examples, the former contains 2,806 images and 188,282 instances, and the latter contains 23,463 images and 192,472 instances. The instance annotation format is the eight-parameter coordinates of the four corner points of the rotation box surrounding the target to be detected. The large number of instances and the complex label format consume a lot of manpower and material resources, prolong the production cycle of open source annotation datasets, and are not conducive to the development and performance improvement of deep learning models.

[0004] To reduce the annotation cost of object detection, many works have explored weakly supervised settings, where the model uses only image-level annotations during training to learn the precise location of the target to be detected. However, whether for ordinary scenes or remote sensing, these models all use weak supervision for horizontal object detection, and there is no precedent for detecting rotated objects using image-level annotations. Given the difficulty of using only image-level annotations, some work has relaxed the restrictions of weak supervision, using horizontal detection boxes as annotations to predict rotated detection boxes. While this weakly supervised setting alleviates some of the annotation burden, labeling horizontal boxes remains cumbersome. Summary of the Invention

[0005] In order to solve the technical problems existing in the background technology, the present invention proposes a method based on visual Figure 1 Weakly supervised remote sensing image rotation target detection method based on consistent network.

[0006] The present invention proposes a Figure 1 The weakly supervised remote sensing image rotation target detection method based on the consistent network includes:

[0007] The original satellite image I is preprocessed to obtain a set of horizontal candidate regions {H1,…,H N};

[0008] The backbone network is used to extract the feature map M∈R of the original satellite image I h×w×l , where h, w, and l are the feature map length, width, and channel dimension respectively;

[0009] The feature map M∈R h×w×l and the horizontal candidate region set {H1,…,H N Input a pre-established and trained rotation target detection model to obtain a first type of rotation detection frame, and use the first type of rotation detection frame as the target recognition result.

[0010] Preferably, the rotation target detection model includes a candidate region alignment module, a rotation angle prediction module, a rotation candidate region alignment module, a K-stage PCL online instance classifier refinement network and a non-maximum suppression module.

[0011] Preferably, the feature map M∈R h×w×l and the horizontal candidate region set {H1,…,H N Inputting a pre-established and trained rotation target detection model to obtain a first type of rotation detection frame, and using the first type of rotation detection frame as the target recognition result, specifically including:

[0012] The feature map M∈R h×w×l and the horizontal candidate region set {H1,…,H N Input the candidate region alignment module to obtain the first feature vector corresponding to each candidate region

[0013] The first eigenvector Input the rotation angle prediction module to obtain the rotation candidate box feature vector and the rotation candidate area set

[0014] The feature map M∈R h×w×l and rotate the candidate region set Input the rotation candidate region alignment module to obtain the second feature vector corresponding to each rotation candidate region

[0015] For the second eigenvector After two linear-ReLu layers mapped to the d3 dimension, the final rotated candidate box feature vector is generated.

[0016] Refine the network and rotate the feature vector of the candidate box based on K-stage PCL online instance classifier To obtain the average value set of instance category scores of all refined branch predictions corresponding to each set of rotation candidate regions

[0017] Rotate the candidate region set Average value collection The preset intersection-over-union ratio threshold is input into the non-maximum suppression module to obtain a first type of rotation detection frame, and the first type of rotation detection frame is used as the target recognition result.

[0018] Preferably, the preprocessing is specifically:

[0019] The original satellite image is segmented by selective search, and then the regions are merged according to the color, texture, scale and filling similarity between regions to obtain a set of horizontal candidate regions {H1,…,H N}.

[0020] Preferably, the feature map M∈R h×w×l and the horizontal candidate region set {H1,…,H N Input the candidate region alignment module to obtain the first feature vector corresponding to each candidate region Specifically include:

[0021] In the feature map M∈R h×w×l Traverse each horizontal candidate region and keep the floating point boundaries unquantized;

[0022] The horizontal candidate region is divided into d1×d1 units, and the boundary of each unit is also kept in floating point form;

[0023] In each unit, the coordinate positions of the four corner points of the candidate box are calculated, and the values ​​of these four coordinate positions are calculated using the bilinear interpolation method. Then, the maximum pooling operation is performed to obtain the first eigenvector corresponding to each candidate area. h, w, and l are the feature map length, width, and channel dimension, respectively.

[0024] Preferably, the first eigenvector Input the rotation angle prediction model to obtain the rotation candidate box feature vector and the rotation candidate area set Specifically include:

[0025] The first eigenvector Input rotation angle prediction fully connected layer Ψ FC-5 , predict the offset and rotation angle of the candidate region corresponding to the feature vector of the candidate region, and obtain the output vector corresponding to the feature vector of the candidate region (t x ,t y ,t w ,th ,t θ );

[0026] The output vector (t x ,t y ,t w ,t h ,t θ ) and the horizontal candidate region set {H1,…,H N} Input the detection box bias decoder to obtain a set of rotated candidate regions

[0027] Preferably, the K-stage PCL online instance classifier refines the network and rotates the candidate box feature vector To obtain the average value set of instance category scores of all refined branch predictions corresponding to each set of rotation candidate regions Specifically include:

[0028] Rotate the candidate box feature vector Input the K-stage PCL online instance classifier refinement network to obtain the candidate region score of each stage

[0029] Obtaining the feature vector of the rotated candidate box The first average of the candidate region scores of the corresponding K stages to obtain each set of rotated candidate regions The average set of instance category scores of all corresponding refined branch predictions

[0030] Preferably, the candidate region set is rotated Average value collection The preset intersection-over-union ratio threshold is input into the non-maximum suppression module to obtain a first type of rotation detection frame, and the first type of rotation detection frame is used as the target recognition result, specifically including:

[0031] Sort the N rotation candidate regions by their corresponding average values, and add the rotation candidate region with the highest current confidence to the final retained result. The current confidence is the average value of the category score.

[0032] Calculating the intersection-and-union ratios (IoU) of the current rotation candidate region and the other rotation candidate regions one by one for the N rotation candidate regions, deleting all rotation candidate regions whose IoU ratios are greater than a preset IoU threshold, and using the remaining rotation candidate regions as the first rotation candidate region set;

[0033] sorting the rotation candidate regions in the first rotation candidate region set according to the average value and then repeating the rotation candidate region retention process until only one rotation candidate region remains;

[0034] The only remaining rotation candidate region is recorded as the first type rotation detection frame, and the first type rotation detection frame is used as the target recognition result.

[0035] In the present invention, the proposed Figure 1 A consistent network-based weakly supervised method for rotated object detection in remote sensing images addresses the lack of rotation angle information in image-level annotations. A neural network architecture is designed around three consistency constraints: image-level labels, rotated detection bounding box positions, and cluster center distributions under different rotations. This method proposes a rotated object detection model based on image-level annotations for remote sensing scenarios. The trained rotated object detection model is then used to process satellite images to obtain the final object recognition results. This reduces the manual annotation workload and improves the efficiency of weakly supervised object detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The present invention proposes a Figure 1 Schematic diagram of the workflow of the weakly supervised remote sensing image rotation target detection method based on consistent network;

[0037] Figure 2 The present invention proposes a Figure 1 Schematic diagram of the processing architecture of the weakly supervised remote sensing image rotation target detection method based on the consistent network;

[0038] Figure 3 The present invention proposes a Figure 1 Schematic diagram of the model training process of the weakly supervised remote sensing image rotation target detection method based on the consistent network;

[0039] Figure 4 The present invention proposes a Figure 1 Schematic diagram of the basic MIL network structure of the weakly supervised remote sensing image rotation target detection method based on the consistent network;

[0040] Figure 5 The present invention proposes a Figure 1 Schematic diagram of the instance classifier refinement branch structure of the weakly supervised remote sensing image rotation target detection method based on consistent network. DETAILED DESCRIPTION

[0041] Reference Figure 1-5 The present invention proposes a method based on visual Figure 1 The weakly supervised remote sensing image rotation target detection method based on the consistent network includes:

[0042] Step S1: pre-process the original satellite image I to obtain a set of horizontal candidate regions {H1,…,H N}.

[0043] In this embodiment, the preprocessing is specifically as follows:

[0044] The original satellite image is segmented by selective search, and then the regions are merged according to the color, texture, scale and filling similarity between regions to obtain a set of horizontal candidate regions {H1,…,H N}.

[0045] Step S2: Use the backbone network to extract the feature map M∈R of the original satellite image I h×w×l , where h, w, and l are the feature map length, width, and channel dimension, respectively.

[0046] Step S3: transform the feature map M∈R h×w×l and the horizontal candidate region set {H1,…,H N Input a pre-established and trained rotation target detection model to obtain a first type of rotation detection frame, and use the first type of rotation detection frame as the target recognition result.

[0047] In this embodiment, the rotation object detection model includes a candidate region alignment module, a rotation angle prediction module, a rotation candidate region alignment module, a K-stage PCL online instance classifier refinement network, and a non-maximum suppression module.

[0048] In this embodiment, if Figures 3 to 5 As shown in Figure 2, the training process of the rotation target detection model is as follows:

[0049] (1) Obtaining original satellite image I (0) , use the Selective Search (SS) algorithm to process it and quickly generate a set of horizontal candidate regions of interest (HRoI) that may contain targets Where N is the number of HRoIs, It can be expressed as (x, y, w, h), which is the horizontal candidate box form of the two-dimensional coordinates of the center point, width and length.

[0050] (2) to I (0) Perform K random rotation transformations to obtain a set of rotation views For H (0) Perform the same rotation transformation to obtain the horizontal candidate area set corresponding to each view where the rotation angle Δθ∈{0,π,π / 2,3π / 2}.

[0051] (3) Using the backbone network to extract the original satellite image I (0) and the collection of all rotated views Where h, w, and l are the feature map length, width, and channel dimension, respectively.

[0052] (4) Use RoI Align to rotate the view M (k) Extract Feature vector of candidate region corresponding to the position Where d1×d1 is the number of sub-regions (bins) divided by RoI Align.

[0053] (5) Fully connected layer Ψ for rotation angle prediction FC-5 by As input, predict the offset and rotation angle of the corresponding rotated region of interest (RRoI). Formally, Ψ FC-5 For features The output vector (t x ,t y ,t w ,t h ,t θ ) has the following regression objectives:

[0054]

[0055] Among them, (x r ,y r ,w r ,h r ,θ r ) indicates based on Generated rotated candidate boxes Center point 2D coordinates, width, length and rotation angle.

[0056] (6) Detection frame bias decoder to And the fully connected layer output vector (t x ,t y ,t w ,t h ,t θ ) is input, and decoded according to the formula in step 4 Five parameters are used to obtain the RRoI sets of K+1 views.

[0057] (7) Using RRoIWarping in M (k) Extracting rotation candidate boxes of different rotation views Feature vector of candidate region corresponding to the position Similar to RoIAlign, where d2×d2 is the number of sub-regions divided by RRoI Warping.

[0058] (8) After further processing by two linear-ReLu layers, Mapped to the d3 dimension. Generate the final rotation candidate box feature vector under K+1 views

[0059] (9) F r In the PCL (Proposal Cluster Learning) online instance classifier refinement network composed of the K-stage collection input, the basic MIL network in the initial stage uses the rotation features generated by the original image As input, generate instance-level pseudo labels under the constraint of image-level classification labels y At the same time Sum the candidate box dimensions to get the image-level prediction label The K instance classifier refinement branches in the optimization phase are respectively based on the rotation features of the K transformed views The instance-level labels predicted in the previous stage are input to generate pseudo labels in sequence as a supervisory signal for the next stage.

[0060] (10) Use the image-level prediction labels generated in step 9 and image-level classification labels y = [y1,…,y C ]Calculate multi-class cross entropy loss Perform image-level classification constraints.

[0061] (11) Use the instance-level pseudo labels generated in each stage of step 9 As supervision, a candidate box level classification loss is generated for each refinement stage The weighted summation can be used to optimize the network parameters at each stage, where W k is the loss weight of the k-th refinement stage.

[0062] (12) Known I (0) To a certain rotation view I (k) The transformation matrix is ​​R Δθ , then the transformation relationship of the rotation candidate region in the RRoI set under K+1 views obtained in step 6 is as follows: The rotation candidate region under the original view The same transformation can be obtained The conversion relationship is as follows:

[0063]

[0064] Among them, (x c ,y c ) represents the rotation center, that is, the center coordinate of the input image. The candidate box corresponding to the k-th rotation view With spatial position and rotation angle consistency, the rotation consistency loss is constructed for each stage / view as follows:

[0065]

[0066] Among them, Ind + is the index set of the positive candidate box, Indicates the confidence of the cluster center of the i-th positive candidate box in stage k, L reg is the regression loss that measures the center point coordinates and angle deviation.

[0067] (13) The inputs of the K refinement stages correspond to K rotated views respectively. The rotation consistency loss is calculated for the positive candidate boxes generated in each stage and the corresponding original image candidate boxes. The final loss is the sum of the losses of each stage:

[0068]

[0069] (14) Use the instance-level pseudo-labels generated by the last refinement branch in step 9 The rotation candidate box set of the original image And the set of rotation candidate boxes of the k-th transformed view Generate cluster centers and where b i ∈R 5 is the rotation box quintuple, p i ∈R C Indicates the class probability.

[0070] (15) Since the actual distribution of the rotation candidate boxes generated by the rotation prediction module under different views cannot be completely consistent, the cluster centers constructed based on spatial similarity also have differences, that is, For this purpose, the cluster center consistency loss L is introduced CC Measure this difference. CC Based on the Hungarian matching loss in the DETR target detection model. First, use (denoting no target) element will With B k The collection is filled to the same size M = max{M0,M k}. Construct matching cost L match It is necessary to consider the category similarity L at the same time cls Similarity L with the spatial position of the detection frame box , using the Hungarian algorithm to find the minimum cost bipartite matching for two sets The specific definitions are as follows:

[0071]

[0072] L cls is the cross entropy loss of the detection box category probability in the two sets. L box Based on KL divergence loss (Kullback-Leibler Divergence), where D kl (·) is used to calculate the KL distance of the rotated detection box converted into a two-dimensional Gaussian distribution, and f(·) performs a nonlinear transformation on the distance to make the loss smoother.

[0073] (16) The final cluster consistency loss is the average cost of all matching pairs under the optimal matching:

[0074]

[0075] (17) Calculate the cluster consistency loss for each of the K rotated views and the original image. The final loss is the sum of the losses of each view:

[0076]

[0077] (18) The final loss of the model is the weighted sum of the image-level category loss, the instance-level loss generated in the PCL refinement stage, and the two consistency constraint losses:

[0078] L=L MIL +ω1L PCL +ω2L RC +ω3L CC ;

[0079] In this embodiment, the training of the rotation target detection model is completed by steps (1) to (18) to obtain a trained rotation target detection model. In order to solve the problem of lack of rotation angle information in image-level annotation, a neural network structure is designed with three consistency constraints of image-level labels, rotation detection box positions and cluster center distribution under different rotation views as the core, and a rotation target detection model based on image-level annotation in remote sensing scenarios is proposed. First, a rotation target detection paradigm is designed based on a typical horizontal target detection model in weak supervision tasks, and the prediction results of the refinement branch are progressively optimized and adjusted by exploring the implicit constraints at the image level; secondly, based on the consistency of the shape and angle of the rotation detection box under different rotation views, the constraint on the predicted angle is realized; finally, since the distribution of the cluster center nodes of the rotation detection box between different rotation views is similar, the Hungarian matching loss is introduced to measure their consistency and guide the optimization direction of the model.

[0080] In this example, rotating objects in remote sensing imagery can be effectively detected using only minimal annotation information. Traditional fully supervised methods rely on complete rotation box annotations, which are manually produced, time-consuming and subjective, and can lead to inconsistent annotation standards across different scenarios. By utilizing only image-level annotations, the proposed model can help reduce the annotation workload, accelerate the annotation process, and alleviate annotation confusion in different situations, thereby enhancing the scalability and applicability of remote sensing object detection in various scenarios and domains.

[0081] In this embodiment, step S3 specifically includes:

[0082] Step S31: transform the feature map M∈R h×w×l and the horizontal candidate region set {H1,…,H N Input the candidate region alignment module to obtain the first feature vector corresponding to each candidate region

[0083] In this embodiment, step S31 specifically includes:

[0084] Step S311: In the feature map M∈R h×w×l Traverse each horizontal candidate region and keep the floating point boundaries unquantized;

[0085] Step S312: Divide the horizontal candidate area into d1×d1 units, and the boundary of each unit also maintains the floating point form;

[0086] Step S313: Calculate the coordinate positions of the four corner points of the candidate box in each unit, use bilinear interpolation to calculate the values ​​of these four coordinate positions, and then perform the maximum pooling operation to obtain the first feature vector corresponding to each candidate area. h, w, and l are the feature map length, width, and channel dimension, respectively.

[0087] Step S32: The first eigenvector Input the rotation angle prediction module to obtain the rotation candidate box feature vector and the rotation candidate area set

[0088] In this embodiment, step S32 specifically includes:

[0089] Step S321: The first eigenvector Input rotation angle prediction fully connected layer Ψ FC-5 , predict the offset and rotation angle of the candidate region corresponding to the feature vector of the candidate region, and obtain the output vector corresponding to the feature vector of the candidate region (t x ,t y ,t w ,th ,t θ ).

[0090] In this embodiment, the output vector (t x ,t y ,t w ,t h ,t θ ) has the following regression objectives:

[0091]

[0092] Among them, (x r ,y r ,w r ,h r ,θ r ) indicates that the n Generated rotation candidate box H n rotate Center point 2D coordinates, width, length and rotation angle.

[0093] Step S322: Output vector (t x ,t y ,t w ,t h ,t θ ) and the horizontal candidate region set {H1,…,H N} Input the detection box bias decoder to obtain a set of rotated candidate regions

[0094] Step S33: transform the feature map M∈R h×w×l and rotate the candidate region set Input the rotation candidate region alignment module to obtain the second feature vector corresponding to each rotation candidate region

[0095] Step S34: the second eigenvector After two linear-ReLu layers mapped to the d3 dimension, the final rotated candidate box feature vector is generated.

[0096] Step S35: Refine the network and rotate the candidate box feature vector based on the K-stage PCL online instance classifier To obtain the average value set of instance category scores of all refined branch predictions corresponding to each set of rotation candidate regions

[0097] In this embodiment, step S35 specifically includes:

[0098] Step S351: rotate the candidate frame feature vector Input the K-stage PCL online instance classifier refinement network to obtain the candidate region score of each stage

[0099] Step S352: Obtain the feature vector of the rotated candidate frame The first average of the candidate region scores of the corresponding K stages to obtain each set of rotated candidate regions The average set of instance category scores of all corresponding refined branch predictions

[0100] Step S36: Rotate the candidate region set Average value collection The preset intersection-over-union ratio threshold is input into the non-maximum suppression module to obtain a first type of rotation detection frame, and the first type of rotation detection frame is used as the target recognition result.

[0101] In this embodiment, step S36 specifically includes:

[0102] Step S361: Sort the N candidate regions for rotation by their corresponding average values, and add the candidate region for rotation with the highest current confidence to the final retained result. The current confidence is the average value of the category scores.

[0103] Step S362: Calculate the IoU ratios of the current rotation candidate region and other rotation candidate regions one by one for the N rotation candidate regions, delete all rotation candidate regions whose IoU ratios are greater than a preset IoU threshold, and use the remaining rotation candidate regions as the first rotation candidate region set;

[0104] Step S363: sort the rotation candidate regions in the first rotation candidate region set according to the average value and then repeat steps S361 and S362 until only one rotation candidate region remains;

[0105] Step S364: record the only remaining rotation candidate region as a first-type rotation detection frame, and use the first-type rotation detection frame as the target recognition result.

[0106] In this embodiment, the proposed Figure 1A consistent network-based weakly supervised method for rotated object detection in remote sensing images employs a self-supervised paradigm, constructing constraints on the rotation angle of the detection bounding box using different rotational views. The entire weakly supervised network architecture follows the PCL (Proposal Cluster Learning) network structure for weakly supervised horizontal object detection. It first transforms the object detection task into a basic MIL (Multiple Instance Learning) problem, aggregating candidate bounding box-level labels and using image-level labels for supervision. Simultaneously, pseudo-labels are generated for the next stage, followed by step-by-step refinement. By constructing a multi-level instance classifier with refined branches, the accuracy of instance-level labels is continuously optimized, aiming to address the problem of weakly supervised object detectors only capturing salient parts of objects. First, a proposal method (ProposalMethod) from a traditional object detection algorithm is used to generate horizontal proposal regions for the image. The image is then randomly rotated to expand the original image into multiple rotated views, and the same rotation transformation is also applied to the horizontal proposal regions. All views are fed into a weighted feature extraction network to generate multiple feature maps. These feature maps and their corresponding transformed proposal bounding boxes serve as input to the rotation angle prediction module, resulting in a proposal vector with the rotation angle. In this case, there is no supervisory signal for the rotation angle, so we need to use consistency conditions across different views to create constraints. Here, we construct three consistency constraints for different rotation views. These three consistency constraints are then used to supervise the application of the rotation object detection model.

[0107] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A weakly supervised remote sensing image rotation target detection method based on view consistency network, characterized by: include: The original satellite image I is preprocessed to obtain a set of horizontal candidate regions {H1,…,H N }, where H N Indicates the Nth horizontal candidate region; N indicates the number of candidate regions; The backbone network is used to extract the feature map M∈R of the original satellite image I h×w×l , where h, w, and l are the feature map length, width, and channel dimension respectively; The feature map M∈R h×w×l and the horizontal candidate region set {H1,…,H N Input a pre-established and trained rotation target detection model to obtain a first type of rotation detection frame, and use the first type of rotation detection frame as the target recognition result; The training process of the rotating target detection model includes: For the original image I (0) Perform K random rotation transformations to obtain a set of rotation views The inputs of the K refinement stages correspond to K rotated views respectively. The rotation consistency loss is calculated for the positive candidate boxes generated in each stage and the corresponding original image candidate boxes. The final loss is the sum of the losses of each stage: in, Among them, Ind + is the index set of the positive candidate box, Indicates the confidence of the cluster center of the i-th positive candidate box in stage k, L reg To measure the regression loss of the center point coordinates and angle deviation; The original image I (0) The rotation candidate region of the local image i(0); is the rotation candidate region corresponding to the local view i(k) of the k-th rotation view; Leveraging instance-level pseudo-labels The rotation candidate box set of the original image And the set of rotation candidate boxes of the k-th transformed view Generate cluster centers and The cluster center distribution difference loss is calculated using the Hungarian matching algorithm; The rotation target detection model includes a candidate region alignment module, a rotation angle prediction module, a rotation candidate region alignment module, a K-stage PCL online instance classifier refinement network and a non-maximum suppression module.

2. The method for detecting rotated objects in remote sensing images based on a weakly supervised view consistency network according to claim 1 is characterized in that: The feature map M∈R h×w×l and the horizontal candidate region set {H1,…,H N Inputting a pre-established and trained rotation target detection model to obtain a first type of rotation detection frame, and using the first type of rotation detection frame as the target recognition result, specifically including: The feature map M∈R h×w×l and the horizontal candidate region set {H1,…,H N Input the candidate region alignment module to obtain the first feature vector corresponding to each candidate region The first eigenvector Input the rotation angle prediction module to obtain the rotation candidate box feature vector and the rotation candidate area set The feature map M∈R h×w×l and rotate the candidate region set Input the rotation candidate region alignment module to obtain the second feature vector corresponding to each rotation candidate region For the second eigenvector After two linear-ReLu layers mapped to the d3 dimension, the final rotated candidate box feature vector is generated. Among them, d1×d1 is the number of units into which the horizontal candidate region is divided; d2×d2 is the number of sub-regions divided by RRoI Warping; Refine the network and rotate the feature vector of the candidate box based on K-stage PCL online instance classifier To obtain the average value set of instance category scores of all refined branch predictions corresponding to each set of rotation candidate regions Rotate the candidate region set Average value collection The preset intersection-over-union ratio threshold is input into the non-maximum suppression module to obtain a first type of rotation detection frame, and the first type of rotation detection frame is used as the target recognition result.

3. The method for detecting rotated objects in remote sensing images based on a weakly supervised view consistency network according to claim 1, wherein: The pre-processing is specifically as follows: The original satellite image is segmented by selective search, and then the regions are merged according to the color, texture, scale and filling similarity between regions to obtain a set of horizontal candidate regions {H1,…,H N }.

4. The method for detecting rotated objects in remote sensing images based on a weakly supervised view consistency network according to claim 2, wherein: The feature map M∈R h×w×l and the horizontal candidate region set {H1,…,H N Input the candidate region alignment module to obtain the first feature vector corresponding to each candidate region Specifically include: In the feature map M∈R h×w×l Traverse each horizontal candidate region and keep the floating point boundaries unquantized; The horizontal candidate region is divided into d1×d1 units, and the boundary of each unit is also kept in floating point form; In each unit, the coordinate positions of the four corner points of the candidate box are calculated, and the values ​​of these four coordinate positions are calculated using the bilinear interpolation method. Then, the maximum pooling operation is performed to obtain the first eigenvector corresponding to each candidate area. h, w, and l are the feature map length, width, and channel dimension, respectively.

5. The method for detecting rotated objects in remote sensing images based on a weakly supervised view consistency network according to claim 2, wherein: The first eigenvector Input the rotation angle prediction model to obtain the rotation candidate box feature vector and the rotation candidate area set Specifically include: The first eigenvector Input rotation angle prediction fully connected layer Ψ FC-5 , predict the offset and rotation angle of the candidate region corresponding to the feature vector of the candidate region, and obtain the output vector corresponding to the feature vector of the candidate region (t x ,t y ,t w ,t h ,t θ ); The output vector (t x ,t y ,t w ,t h ,t θ ) and the horizontal candidate region set {H1,…,H N } Input the detection box bias decoder to obtain a set of rotated candidate regions 6. The method for detecting rotated objects in remote sensing images based on a weakly supervised view consistency network according to claim 2, wherein: The K-stage PCL online instance classifier refines the network and rotates the candidate box feature vector To obtain the average value set of instance category scores of all refined branch predictions corresponding to each set of rotation candidate regions Specifically include: Rotate the candidate box feature vector Input the K-stage PCL online instance classifier refinement network to obtain the candidate region score of each stage Obtaining the feature vector of the rotated candidate box The first average of the candidate region scores of the corresponding K stages to obtain each set of rotated candidate regions The average set of instance category scores of all corresponding refined branch predictions 7. The method for detecting rotated objects in remote sensing images based on a weakly supervised view consistency network according to claim 6, wherein: The candidate region set will be rotated Average value collection The preset intersection-over-union ratio threshold is input into the non-maximum suppression module to obtain a first type of rotation detection frame, and the first type of rotation detection frame is used as the target recognition result, specifically including: Sort the N rotation candidate regions according to their corresponding average values, and add the rotation candidate region with the largest average value to the final retained result; Calculating the intersection-and-union ratios (IoU) of the current rotation candidate region and the other rotation candidate regions one by one for the N rotation candidate regions, deleting all rotation candidate regions whose IoU ratios are greater than a preset IoU threshold, and using the remaining rotation candidate regions as the first rotation candidate region set; sorting the rotation candidate regions in the first rotation candidate region set according to the average value and then repeating the rotation candidate region retention process until only one rotation candidate region remains; The only remaining rotation candidate region is recorded as the first type rotation detection frame, and the first type rotation detection frame is used as the target recognition result.

Images